Re:Heretic: Automatic censorship removal for language models
by fer on Monday November 17, 2025 @03:46AM
(#45951899)
Haven't tested the latest DeepSeek versions, but the first release wasn't censored as a model on Taiwan. The issue is that if you use their app (as opposed to locally), it replaces the ongoing response with "sorry can't help" once it starts saying things contrary to the CCP dogma.
[ Parent
| on HN ]
Re:Heretic: Automatic censorship removal for language models
by kstrauser on Monday November 17, 2025 @07:52AM
(#45953135)
I ran it locally and it flat-out refused to discuss Tiananmen Square ‘88. The “thinking” clauses would display rationales like “the user is asking questions about sensitive political situations and I can’t answer that”. Here’s a copy and paste of the exact conversation: https://honeypot.net/2025/01/27/i-like-running-ollama-on.htm...